Classification of Alzheimer s disease subjects from MRI using the principle of consensus segmentation Aymen Khlif and Max Mignotte 1 st September, Maynooth University, Ireland
Plan Introduction Contributions : Proposed classification model Dataset description MRI data preprocessing Prototypes NC and AD Hybrid classification Experiments and results Conclusion and perspectives 2
Introduction : Alzheimer's disease (AD) Neurodegenerative and progressive disease Deterioration of cognitive functions Memory, language and behavior disorders Progressive loss of autonomy Disorientation in time and space 3
Introduction : Statistics of Alzheimer s Disease 1 In 2050: 115 million in the world, a new case every 7 seconds 13.5 million in the USA (245454 new cases per year) Global economic impact: $ 600 billion in 2010 Urgent issue: Early diagnosis of AD Implementation of a therapeutic Slow down the neurodegenerative process 1 Association Alzheimer s disease International 4
Introduction : Magnetic Resonance Imaging (MRI) Brain analysis: structural, functional and non-invasive Contributing to the early diagnosis of AD Structural alterations Metabolic alterations Detecting changes Macro-structural Micro-structural IRM Anatomic/Structural Diffusion Tensor Imaging 5
AD diagnosis methods Volumetric analysis / Area of interest Study the variation of the volume of a region Manual: time-consuming, depends on the observer (clinician) Automatic / semi-automatic: suffers from errors voxel analysis / Voxel To detect significant differences in Grey Matter (GM) between two groups of subjects by voxel-to-voxel tests Do not require a priori assumptions about the location, the size or number of ROIs to be analyzed, since they provide voxel wise measures determined in the entire brain Help to detect structural changes in MRIs Do not depend on the clinician abilities Methods for group analysis: Individual diagnosis? 6
Problematic Individual diagnosis = visual assessment of a new case Learn about similar cases? Detection and characterization of pathological targets? To which class of known subjects can it be associated? Lists of similar images? 7
Solution Use recent advances made in segmentation and multimedia Indexing 2 and classification for Content Based Visual Information Retrieval (CBVIR). More precisely, use the concept of consensus segmentation to build two segmentation prototypes (Prototype Normal Control and Prototype Alzheimer's Disease) Tools Indexing by visual content of images Principle of consensus segmentation based atlas Using Domain Knowledge" in: Image acquisition MRI Diagnosis of AD 2 A.Khlif and M.Mignotte (2017). Visualisation et mise en cluster des données de segmentation. Outils et applications multimédias, 76 (1), 1531-1552. 8
principle of consensus segmentation A consensus segmentation is conceptually the compromise (in terms of level of details, contour accuracy, number of regions, etc.) exhibited by each segmentation map (or spatial clustering) belonging to a set of segmentations In our case, the principle of consensus segmentation allows us to build two reliable segmentation-based prototypes, one corresponding to healthy individuals and the second one corresponding to unhealthy subjects (with AD) The segmentation into three kinds of regions has the merit to efficiently reduce the information content of a brain image and to suppress noise and artifacts which are not relevant for the AD detection These two consensus segmentation-based prototypes allow us to suppress undesired components in the brain image (to be classified) such as the anatomical variability existing between individuals which are not relevant for the detection and quantification of AD 9
Data description OASIS (Open Access Series of Imaging Studies: www.oasis-brains.org ) Worldwide project Sharing data for research in the treatment and diagnosis of Alzheimer's disease We will consider a subset of the complete cross-sectional OASIS dataset, with 49 controls and 49 AD patients 10
MRI data preprocessing Affine registration to template Brain masking (BET) Grey/White/CSF segmentation Talairach atlas brain 11
Construction of prototypes 12
Proposed classification model Pott distances is the normalized number of labels differences in percentage 13
Comparison with morphometric methods (NC vs. AD) 14
Conclusion Our approach is automatic and does not require the intervention of the clinician during the disease diagnosis It is extensible to other diseases that can be diagnosed by brain MRI such as Schizophrenia and brain tumors Perspectives The method could be extended by combining axial, coronal, and sagittal MRI data for improving the classification accuracy Generalize the approach for the 3D case and compare it with 2D Classification in four classes (NC, Very mild AD, mild AD, moderate AD) Generalize the approach for other criteria for the consensus segmentation (e.g., VOI, GCE, PRI, FCR ) 15
Perspectives 16
17